Anthropic's Claude Sonnet 5.5 is 30% faster at the same price
Anthropic says its new mid-tier model runs 30% faster and costs up to 30% less per task than Sonnet 5, and is the first Sonnet shipped with Opus-class cyber safeguards.
Anthropic has released Claude Sonnet 5.5, a mid-tier model that the company says runs more than 30% faster than Sonnet 5 while costing up to 30% less per task. It arrived on 28 September as the second member of the Claude 5.5 family, a week after the flagship Opus 5.5, and it is the first Sonnet model Anthropic has shipped with the cybersecurity safeguards it usually reserves for its most capable systems.
The pricing is unchanged, which is the interesting part. Sonnet 5.5 costs $2 per million input tokens, $10 per million output tokens and $0.20 per million cache reads — identical to Sonnet 5. Anthropic's claim is that the cheaper number comes from doing the same work with far fewer tokens rather than from a discount, because the model batches tool calls and takes fewer steps. On the company's own benchmark chart, Sonnet 5.5 at low or medium effort beats Sonnet 5's best score for about a tenth of the cost per task.
What the numbers actually say
Anthropic reports 70.6% on Terminal-Bench 4.0, an agentic coding evaluation, against Sonnet 5's 10.3%, and says the model lands two points below Opus 5.5 on GDPval-AA, which tests real-world occupational work. It also notes, fairly, that Sonnet 5.5 at maximum effort can perform comparably to Opus 5.5 on several evaluations while still trailing it on long, open-ended problems that need sustained judgment. There is a genuine claim buried in the footnote: Sonnet 5.5 is the first Sonnet model to beat Pokémon Red working from screenshots alone, which is a crude test, but a real one for image understanding and long-horizon play.
Because Anthropic rates its cyber capabilities as comparable to Opus 5's, Sonnet 5.5 launches with the cyber safeguards and fallbacks normally attached to the frontier models, and Anthropic says it is the first Sonnet with classifiers that block reasoning extraction — the distillation attacks where someone rents thousands of accounts to copy a model's behaviour. Biology safeguards stay as they were on Sonnet 5. The company says the models are available now on Amazon Web Services, Google Cloud and Microsoft Azure, with Claude Haiku 5.5 promised in the coming weeks.
A mid-tier model with frontier manners
The pattern across the 5.5 family is that the cheap seat keeps getting closer to the expensive one. Anthropic's own framing is that Opus 5.5 remains the model for hard, ambiguous work and Sonnet 5.5 is for well-scoped tasks, bug fixes, documents, slides and spreadsheets. In practice, most people's work is the second category, which is why the price-per-task figure matters far more than the leaderboard.
Our opinion
Every model launch now comes with the same ritual: the vendor's benchmarks decide the headline, and independent testers spend the following fortnight finding out what they missed. So treat the 70.6% as a claim, not a fact. What is not in dispute is the shape of the release. Anthropic has put the more capable, better-guarded version of its workhorse model on the same price list as the old one and let the token bill do the talking — which is a quieter, more useful kind of progress than another record score. The cyber safeguards arriving on a mid-tier model is the detail worth noting, because it says the company now expects cheap models to be dangerous in the same way expensive ones are. If Haiku 5.5 follows on the same curve, the argument over whether capable AI is too expensive will have moved on. It will simply be everywhere.